Such assessments are Response Theory (IRT) to capture temporal dynamics in referred to as formative assessments and are used not only learner ability. While these models have the potential to allow to track student learning and make appropriate instructional instructional systems to actively monitor the evolution interventions, but also to allow learners to practice their of learner proficiency in real time, existing dynamic item response knowledge and skills, and make necessary self-corrections models rely on expensive inference algorithms that [17]. When learning occurs alongside assessment, learner scale poorly to massive datasets. In this work, we propose proficiency is longitudinal rather than inert, and the assumption Variational Temporal IRT (VTIRT) for fast and accurate of static proficiency makes standard IRT less suitable inference of dynamic learner proficiency. VTIRT offers orders as a model of proficiency measurement. of magnitude speedup in inference runtime while still providing accurate inference. Moreover, the proposed algorithm Dynamic Item Response models [14, 11] mitigate this issue is intrinsically interpretable by virtue of its modular by removing the assumptions of static ability and instead allowing design. When applied to 9 real student datasets, VTIRT it to stochastically change over time, but existing inference consistently yields improvements in predicting future learner methods rely on expensive iterative algorithms with performance over other learner proficiency models.